Delivery field notes
How does three-point estimation work in a project?
Three-point estimation with a worked schedule: the PERT formula, what the spread tells you, how to add tasks up honestly and where the method misleads.
One number hides what you know
Ask someone how long a task will take and you get one number. That number blends three things: how long it takes when things go well, what usually happens, and how bad it can get. The listener cannot tell which of the three they heard. Often the estimator cannot either.
Three-point estimation separates them. For each task you give an optimistic estimate, a most likely estimate and a pessimistic estimate. The method dates from the late 1950s, when the US Navy's Polaris missile programme developed PERT to evaluate progress on a large, uncertain development effort. The original paper, by Malcolm, Roseboom, Clark and Fazar in Operations Research in 1959, describes a programme modelled as a network of events, with elapsed-time estimates for the activities between them. The paper expands PERT as Program Evaluation Research Task, the name of the project that built it; the technique later became known as the Program Evaluation and Review Technique.
The formula
For each task you estimate:
- O, optimistic: the time if most things go right. Not a miracle, but a good run.
- M, most likely: the time you would bet on if forced to pick one number.
- P, pessimistic: the time if the realistic problems happen. Not a disaster that would cancel the project.
PERT then uses two calculations:
Expected time E = (O + 4M + P) / 6 Spread SD = (P − O) / 6
The expected time leans towards the most likely value but is pulled up when the pessimistic tail is long. The spread is a rough measure of how uncertain the task is. A task estimated at 4, 5 and 6 days and one estimated at 3, 5 and 15 days have the same most likely value. They are very different tasks to plan around.
A worked example
The project below is fictional: configuring a new access-control system for an office before staff move in. Five tasks run one after another.
| Task | O (days) | M (days) | P (days) | E = (O+4M+P)/6 | SD = (P−O)/6 |
|---|---|---|---|---|---|
| Confirm door specifications | 2 | 3 | 5 | 3.2 | 0.5 |
| Supplier delivers readers | 5 | 8 | 20 | 9.5 | 2.5 |
| Install readers | 3 | 4 | 7 | 4.3 | 0.7 |
| Configure access groups | 2 | 3 | 6 | 3.3 | 0.7 |
| Test with staff cards | 1 | 2 | 4 | 2.2 | 0.5 |
| Total | 13 | 20 | 42 | 22.5 | see below |
Three things stand out.
The most likely total is optimistic. Adding the most likely values gives 20 days. Adding the expected values gives 22.5 days. The difference comes almost entirely from supplier delivery, where the pessimistic case is far longer than the most likely one. Most plans are built on most likely values, which is one reason most plans run late.
One task carries the uncertainty. Supplier delivery has a spread of 2.5 days; the other tasks have 0.5 to 0.7. If you want to make the schedule more reliable, that is where effort pays: a confirmed delivery date, a second supplier, or rental readers as a fallback.
Spreads do not add up directly. For tasks in sequence that vary independently, PERT combines spreads by adding the squares and taking the square root:
Combined SD = √(0.5² + 2.5² + 0.7² + 0.7² + 0.5²)
= √(0.25 + 6.25 + 0.49 + 0.49 + 0.25)
= √7.73 ≈ 2.8 daysAs a rough guide, a date of E plus one combined spread (22.5 + 2.8, so about 25 days) is a more defensible commitment than 22.5, and E plus two spreads (about 28 days) is a cautious one. These ranges assume the tasks vary independently, which the next section questions.
Where the method misleads
Three-point estimation is a big improvement on single numbers, but it has well-known weaknesses.
Pessimistic estimates are rarely pessimistic enough. People anchor on the most likely value and add a modest margin. Bent Flyvbjerg's review of very large projects describes an "iron law": over budget, over time, over and over again. Your project is smaller than a megaproject, but the bias is the same. A useful check: ask, "What happened on the worst similar task we have done?" and use that as P.
Tasks are not independent. If the supplier is late, installation and testing often slip too, because the installer has moved to another job. When delays come from a shared cause, the combined spread is larger than the square-root formula suggests. Treat the formula as a lower bound on uncertainty, not a precise figure.
Merge points add delay. When two paths must both finish before a third task can start, the start waits for whichever path is later. PERT's arithmetic follows one path at a time and understates this effect. If your schedule has parallel paths that join, look at each path and plan around the slower one.
It does not replace a dependency plan. The numbers are only as good as the task list. A missing task, such as "obtain building-owner approval for drilling", adds more risk than any spread calculation captures.
How to run an estimating session
- List tasks small enough to estimate. If someone cannot picture doing the task, split it.
- Ask for P first. Starting with the pessimistic case reduces anchoring on the most likely value. Ask what would cause it.
- Write down the cause of each long tail. "Supplier backlog in November" is more useful than "20 days". The cause tells you what to watch and what fallback to prepare.
- Compute E and SD, then look at the spreads. The task with the largest spread deserves a risk entry with a trigger date.
- Commit to a range. Report "22 to 28 working days, most likely about 25", with the main reason for the range. Then track the high-spread tasks closely.
Turning the range into a commitment
A range is only useful if someone acts on it. When a sponsor asks for a date, give them three things: the date you will commit to, the most likely date, and the one task whose outcome moves it most. "We commit to 28 working days. Most likely is about 25. The supplier delivery is the swing factor, and we will know by the 14th whether it holds."
That answer does two jobs. It gives the sponsor a date they can plan around, and it tells them when to expect better information. If the supplier confirms early, you can bring the date in. If not, the 14th is the day you raise it, not the week before go-live.
Resist the request to "just give one number". A single date with no range invites the sponsor to treat it as certain and treats every slip as a failure. A committed date with a stated swing factor turns a slip into something everyone saw coming and had a plan for.
Try it on a small piece of work
Pick five tasks from a project you know, or the fictional move above with your own numbers, and spend fifteen minutes:
- Write O, M and P for each, asking for P first and noting its cause.
- Calculate E and SD per task, and the totals.
- Compare the sum of M with the sum of E. Where does the difference come from?
- Write the date you would commit to, and the one task you would watch.
The project schedule example shows how to present dependencies alongside dates, and the recovery plan example covers what to do when the high-spread task does slip.
To practise finding the dependency that controls a finish date, try the open exercise Community clinic relocation: project schedule.
Sources
Method background
These sources provide general background. The worked examples and worksheets are original Response Shift teaching material; the organisations do not endorse this product.
